MNIST DIGITS CLASSIFICATION EXAMPLE

This example uses the mnist model to identify a digit from an image. It steps through a simple DKube workflow.

Note This example runs on DKube V3.x and above

1. Setup Resources

Before using DKube to experiment, train, and deploy, the resources must be set up.

Create Code Repo

• From the Code menu on the left, select + Add Code with the following fields:
  • Name: mnist (Or choose <your-code-repo>)
  • Code Source: Git
  • URL:: https://github.com/oneconvergence/dkube-examples.git
  • Branch: tensorflow
• Leave the other fields in their current selection and Add Code

Create Dataset Repo

• From the Datasets menu, select + Add Dataset with the following fields:
  • Name: mnist (Or choose <your dataset-repo)
  • Dataset source: Other
  • URL: https://s3.amazonaws.com/img-datasets/mnist.pkl.gz
• Leave the other fields in their current selection and Add Dataset

Create Model Repo

• From the Models menu, select + Add Model with the following fields:
  • Name: mnist (Or choose <your model-repo)
• Leave the other fields in their current selection and Add Model

2. Create & Launch JupyterLab Notebook

JupyterLab can be used to experiment with your code.

• Ensure that all of Repos above are in the Ready state
• From the IDEs menu, select + Add JupyterLab with the following fields:
  • Basic tab
    • Name: <your-IDE-name (Your choice)
    • Code: <your-code-repo> (Created during the Code Repo step)
    • Framework: tensorflow
    • Framework Version: 2.0.0
    • Image: ocdr/dkube-datascience-tf-cpu-multiuser:v2.0.0-17

    Note The default Tensorflow Image should fill in automatically, but ensure that it is correct
    
    

  • Repos tab
    • Inputs > Datasets: <your-dataset-repo> (Created during the Dataset Repo step)
      • Mount Path: /mnist
• Leave the other fields in their current selection and Submit
  
  
• Once the IDE is running and the JupyterLab icon on the right is active, select it to launch a JupyterLab window
  • Navigate to workspace/<your-code-repo>/mnist
  • Open train.ipynb
    • Run All Cells from the menu at the top
    • Change the EPOCHS variable in the 2nd cell "5" and rerun all cells
    • You can view the difference in output at the bottom of the script

    Note You would normally be developing your code in JupyterLab, and once you were satisfied you would create a Python file from the ipynb file. In this example, a Python file is already ready for execution.

3. Work with Training Runs

Batch training runs can be used to create trained models.

Run Training Job

• From the Runs menu, select + Run > Training with the following fields:
  • Basic tab
    • Name: <your-run-name (Your choice)
    • Code: <your-code-repo> (Created during the Code Repo step)
    • Framework: tensorflow
    • Framework Version: 2.0.0
    • Image: ocdr/dkube-datascience-tf-cpu-multiuser:v2.0.0-17

    Note The default Tensorflow Image should fill in automatically, but ensure that it is correct
    
    

    • Start-up Command: python mnist/train.py
  • Repos tab
    • Inputs > Datasets: <your-dataset-repo> (Created during the Dataset Repo step)
      • Mount Path: /mnist
        
        
    • Outputs > Models: <your-model-repo> (Created during the Model Repo step)
      • Mount Path: /model

    Note Ensure that you add the Model into the Outputs section, and not the Inputs section

  • Leave the other fields in their current selection and Submit
  • Your Run will show up from the Runs menu screen
    
    
  • Clone the Run by selecting the checkbox and choosing Clone from the top buttons
    • Leave the Basic and Repos tabs the same
    • On the Configuration tab
      • Select the + button next to Environment Variables
      • Key: EPOCHS (Must be in upper case)
      • Value: 5
    • Submit

Compare Runs

• Wait for both Runs to complete
• From the Runs menu, select both Run checkboxes, then select Compare button
• Scroll down and choose Y-Axis: train_accuracy

Run Katib-Based Hyperparameter Tuning

• Go to https://github.com/oneconvergence/dkube-examples/tree/tensorflow/mnist/tuning.yaml
• Select Raw
• Right-click & Save as... "tuning.yaml"
  
  
• From the Runs menu, select the first Run checkbox, then select Clone
  • Leave the Basic and Repos tabs the same
  • On the Configuration tab
    • Select Upload Tuning Definition
    • Choose the tuning.yaml file that you saved
  • Submit
    
    
• Wait for Run to complete
• View the results by selecting the Katib icon on the right of the Run line

Tuning.yaml File Details

• objective: The metric that you want to optimize
• goal parameter is mandatory in tuning.yaml file
• objectiveMetricName: Katib uses the objectiveMetricName and additionalMetricNames to monitor how the hyperparameters work with the model. Katib records the value of the best objectiveMetricName metric.
• parameters : The range of the hyperparameters or other parameters that you want to tune for your machine learning (ML) model
• parallelTrialCount: The maximum number of hyperparameter sets that Katib should train in parallel. The default value is 3.
• maxTrialCount: The maximum number of trials to run
• maxFailedTrialCount: The maximum number of failed trials before Katib should stop the experiment
• algorithm: Search algorithm to find the best hyper parameters. Value must be one of following:
  • random
  • bayesianoptimization
  • hyperband
  • cmaes
  • enas

4. Deploy Model

After the best model is identified, it can be deployed for inference serving.

• From Models menu, select <your-model-repo> (Created during Model Repo step)
• Choose the highest version of the Model
• Select the Lineage tab
  • This provides information on the inputs and outputs of the Model
    
    
• Select the Metrics tab
  • This provides the metrics associated with the Model
    
    
• Go back to Models top menu, and reselect the Model
• Select the Deploy icon on the right of the newest Model
  • Name: <your-deploy-name> (Your choice)
  • Deployment: Production
  • Deploy Using: CPU
  • Transformer: Check Box
    • Transformer Script: mnist/transformer.py
  • Leave the other fields in their current selection an Submit
    
    
  • The deployed Model will appear in the Deployments menu screen

5. Train & Deploy with Kubeflow Pipelines

The training and deployment steps can be automated using Kubeflow Pipelines.

• Open the JupyterLab window
• Navigate to workspace/<your-code-repo>/mnist
• Open pipeline.ipynb
  
  
• If you chose the default value for all of your repos (mnist) then Run all Cells
  
  
• If you chose different repo names
  • In the 2nd cell, labeled User Variables, modify the repo names with your chosen names
  • Run All Cells from the menu at the top
    
    
• From the Pipelines menu on the left
  • Select Runs tab
  • Your new pipeline will be executing
  • Select the pipeline name to see its progress

6. Test inference

• Create a browser tab and go to https://<dkube_url>/inference
• Paste the Endpoint URL from Deployments
• Copy Auth token from Developer settings in DKube page and paste in
• Choose mnist for model type
• Download 3.png from repo
• Click Predict

Note The prediction may time out waiting for the pod to start - select wait if prompted